Update: The paper “Active Management of Timing Guardband to Save Energy in POWER7” also earned 2011 Pat Goldberg Memorial Best Paper recognition.
Researchers at IBM’s Austin
lab recently earned “Best Paper” at The 44th Annual IEEE/ACM International Symposium on Microarchitecture for work on “Active Management of Timing Guardband to Save Energy in POWER7” – a new way of managing and reducing
microprocessor power consumption.
Charles Lefurgy, a Master Inventor on the team, discusses how this technology works, and how it might affect chips in everything from servers to smart phones.
How much energy do microprocessors consume in a system –
from mobile devices to mainframes?
Charles Lefurgy: High performance microprocessors
can consume over 200 Watts, while mobile device microprocessors found in smart
phones may only consume about a half of a Watt.
How is this consumption monitored?
CL: In IBM
POWER7 processor-based systems, a dedicated “EnergyScale” microcontroller measures the power
supplied to individual components, such as the microprocessors, disks, memory,
and fans. While smart phones and PCs may have dozens of sensors to monitor
their operating environments, a POWER7-based system has thousands of physical and
virtual sensors to monitor its operation.
For this work, we use a new sensor built
into the POWER7 chip called a “critical-path monitor.” It tells us when the
chip is experiencing overly conservative conditions (high voltage or low
temperature). This knowledge allows us to safely reduce the chip voltage and
reduce chip power consumption by about 20 percent. Another key sensor we added in
POWER7 monitors chip-level performance in order to guarantee that optimizing
for high energy-efficiency does not harm performance.
Talk a little bit about the history of microprocessor
energy usage, and energy saving breakthroughs.
CL: Technology changes associated with
ever smaller CMOS transistor sizes have had a
major impact on reducing the energy used per computation. From a systems
viewpoint, dynamic voltage and frequency scaling has been commercially
available for over a decade, and is used in everything from smart phones to
supercomputers. This allows the power to reduce by lowering the supply voltage
and clock frequency used by the microprocessor when the processing demand
falls.
We have also seen many micro-architecture
improvements in how instructions are processed within the microprocessor. For
example, speculating the outcome of computations so that the next one can begin
before the previous one ends. This reduces the time that the microprocessor
sits idle (using power) while waiting for data to become available for
processing.
How does your use of the “Critical Path Monitors” technology
further improve energy savings in microprocessors?
CL: Today, the voltage level used to
operate a microprocessor is overly conservative. The reason for this is that a
large amount of engineering margin in the form of additional voltage is
used to protect the microprocessor from worst-case conditions that may occur in
the real world. For example, data centers may overheat or there may be an
unexpected workload with high power consumption – both of which may cause the
chip’s voltage level to droop. If the voltage becomes too low, electric signals
cannot propagate through the chip, and it fails.
The new capability we’ve demonstrated
safely removes some of this voltage margin. Operating at a lower voltage level
reduces the chip’s power consumption – in addition to using traditional
techniques.
The key to our work is to use the
Critical Path Monitor’s (CPM) precision sensing of the time it takes for the
circuits within the chip to complete a computation. If that timing changes due
to a voltage droop, then the chip can react very quickly to protect itself.
The protection is enabled by using
a new element in POWER7, called a digital phase-lock loop (DPLL). It continuously
watches the CPM and reacts within nanoseconds to temporarily slow down the chip
clock frequency to operate at a reduced voltage, and avoid a failure. Once we
had this safety mechanism in place, we modified the EnergyScale microcontroller
to look for opportunities to lower a chip’s voltage when it experienced
excessive voltage margins. Voltage is lowered until the DPLL just begins to
lower the frequency that keeps the chip operating correctly.
An additional benefit of using a
lower voltage is that the microprocessor is cooler. This allows the server fans
to run quieter, at a reduced speed and power consumption.
Can it be applied to microprocessors across all types of
systems and devices?
CL: Every microprocessor applies
additional voltage margin to eliminate harmful voltage droops. I expect that
any system from mobile devices to supercomputers could use our method to save
energy.
Is this a redesign that will take time to realize or
something that can be implemented in today's systems and devices?
CL: Our method requires new sensors,
and the DPLL clocking, to be built into the microprocessor. Today, only the
POWER7 processor has this capability.
How much energy savings could be realized – and would it
change system performance?
CL: Our results on an IBM Power 755 system
achieved a power reduction of about 20 percent for the microprocessor and 18
percent for the entire server.
One way this could be used in
practice is to increase the server density in data centers. Data centers often
have power limitations for the amount of equipment that can be installed in a
rack. Therefore, installing these energy-saving servers could allow a data
center to install 18 percent more servers within the same power budget, and
achieve 18 percent higher performance.
We observed no change in performance for industry-standard server
benchmarks. Our control algorithms
adjust voltage rapidly enough to tightly achieve a desired level of
performance.
What would this energy savings mean for a mobile device
or other consumer products?
CL: For mobile devices, the battery
life could be extended at today's performance level. However, the
microprocessor is a small part of the energy consumption of a mobile device
compared to the display. Therefore, the energy savings would be less, percentage-wise,
compared to servers.
Another benefit is that the microprocessor requires less cooling due to running
at a lower voltage. You could imagine using this in game consoles, laptops, and
PCs where you don’t want loud fan noise.
When and in what industry might we see this technology
first utilized?
CL: Critical Path Monitors are actively
used today in IBM Power 775 supercomputers to detect a loss of voltage level. But
currently, rather than reducing the clock frequency, a different method to
protect the system is used – and is the subject of a pending conference
publication.
We expect that our on demand
reduction of microprocessor voltage margin, using CPMs and DPLLs, will appear
in future IBM Power Systems starting in 2013.
How many patents were granted for CPM?
CL: IBM
holds 11 granted patents related to CPM, including the circuitry used to
synthesize a critical path and the circuitry to finely control the clock
frequency output by the DPLL. Another seven patent applications have been filed.
They’re for things such as methods to select the chip voltage level using the
CPM and DPLL.
The co-authors on the research paper are IBMers Alan Drake, Michael
Floyd, Malcolm Allen-Ware, Bishop Brock, Jose Tierno, and John Carter.
Labels: energy, microprocessor, power7